Underground pipe gallery fire feature correlation analysis system based on improved Apriori algorithm

Abstract

The invention relates to the technical field of pipe gallery fire analysis, in particular to an underground pipe gallery fire feature correlation analysis system based on an improved Apriori algorithm, and aims to comprehensively collect underground pipe gallery data through multi-modal sensing equipment and pre-process the underground pipe gallery data into a standardized three-dimensional time sequence data set to provide a high-quality data basis for subsequent analysis; a multi-dimensional feature association matrix is constructed based on the three-dimensional time sequence data set, feature association strength is quantified, and a multi-dimensional feature association basis is provided for fire risk mining; a fire risk rule knowledge base is constructed through a space-time weighted support degree optimization Apriori algorithm, effective association rules are accurately extracted, and the accuracy and reliability of fire risk early warning are improved; based on the fire risk rule knowledge base, a decision tree model is established, association rule weights of branches of the decision tree are analyzed, fire risk grades are divided, a composite early warning instruction set is generated, different coping strategies can be adopted for risks of different grades, and the pertinence and effectiveness of emergency response are improved.

Description

Technical Field

[0001] This invention relates to the field of underground utility tunnel fire analysis technology, and in particular to an underground utility tunnel fire feature correlation analysis system based on an improved Apriori algorithm. Background Technology

[0002] Current fire analysis in underground utility tunnels still faces numerous challenges: Fire monitoring data often originates from multiple sensors and devices, which may be scattered across different systems or platforms, creating data silos. Furthermore, differences in data formats and sampling frequencies among different devices make direct integration and analysis difficult. Traditional fire analysis may focus only on single characteristics, neglecting the inherent connections and interactions between features, potentially leading to misjudgments or omissions in fire risk assessment. Traditional fire analysis may ignore the impact of time and space factors on fire risk, resulting in untimely or inaccurate warnings. The occurrence and development of fires are often closely related to time and space factors; however, traditional fire information analysis methods may not fully consider these factors, leading to inaccurate or untimely warnings. Currently, fire warning systems still employ static warning rules, which are based on historical data or experience and are ill-suited to adapting to changes in the underground utility tunnel environment and the dynamic nature of fire risks. Static rules may also cause the warning system to react slowly to emerging fire patterns or risk factors. Therefore, this invention proposes an underground utility tunnel fire feature correlation analysis system based on an improved Apriori algorithm. Summary of the Invention

[0003] The purpose of this invention is to solve the problems in the background art by proposing an underground utility tunnel fire feature association analysis system based on an improved Apriori algorithm.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm includes: a multimodal data acquisition and preprocessing module, a multidimensional feature association matrix construction module, a dynamic rule base construction and optimization module, and a hierarchical risk early warning output module; Multimodal data acquisition and preprocessing module: Collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices, and generates standardized three-dimensional time-series datasets after data preprocessing. Multidimensional feature association matrix construction module: Based on a three-dimensional time series dataset, statistical features, time-varying features, and spatial features are extracted through a sliding window spatiotemporal slice. Strongly correlated feature pairs are selected by combining the Pearson correlation coefficient, and the correlation strength of feature pairs is quantified by dynamic weights to construct a multidimensional feature association matrix. Dynamic rule base construction and optimization module: Based on the multidimensional feature association matrix, candidate itemsets are generated through data transformation and initial pruning. Then, frequent itemsets are screened by combining spatiotemporal weighted support, and a multi-stage pruning optimization strategy is implemented to build a fire risk rule knowledge base. The graded risk warning output module establishes a decision tree model based on the fire risk rule knowledge base, analyzes the association rule weights of the decision tree branches, classifies fire risk levels, and generates a composite warning instruction set.

[0005] Furthermore, in the multimodal data acquisition and preprocessing module, the multimodal sensing devices include a distributed fiber optic temperature sensor, a gas sensor array, and a smoke concentration detection unit.

[0006] Furthermore, the multimodal data acquisition and preprocessing module collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices. The process of generating a standardized three-dimensional time-series dataset through data preprocessing includes: Distributed fiber optic temperature sensors, gas sensor arrays, and smoke concentration detection units are deployed in the underground utility tunnel; multi-node data synchronous acquisition is achieved using the CAN bus protocol to generate a multi-source heterogeneous data stream of the underground utility tunnel containing temperature, gas concentration, and smoke density. This data stream is the raw data stream. Wavelet transform is applied to the raw data stream to eliminate high-frequency noise, moving average filtering is used to suppress low-frequency drift, and Z-score normalization is used to unify the dimensions of multi-source sensor data; based on GPS timing system and UWB positioning technology, each set of sensor data is assigned a timestamp and spatial coordinates. Construct a spatiotemporal index matrix containing timestamps and spatial coordinates: Treat each sensor data as an element of the matrix, arrange them according to time order and spatial location to form a matrix structure, where the rows of the matrix represent the time dimension and the columns represent the spatial dimension. The matrix elements contain sensor data at the corresponding time point and spatial location, thus forming a standardized three-dimensional time-series dataset.

[0007] Furthermore, the multidimensional feature association matrix construction module, based on a three-dimensional time-series dataset, extracts statistical features, time-varying features, and spatial features through a sliding window spatiotemporal slice. It then uses the Pearson correlation coefficient to screen strongly correlated feature pairs and utilizes dynamic weights to quantify the association strength of these feature pairs. The process of constructing the multidimensional feature association matrix includes: In the constructed three-dimensional time series dataset, based on the spatial layout of the underground utility tunnel, data from different spatial locations at the same timestamp are integrated: sensor data on the cross-section of the utility tunnel are classified and summarized according to their spatial coordinates; by traversing all time points in the three-dimensional time series dataset, the full cross-sectional data of the utility tunnel corresponding to each time point is generated in sequence, thus obtaining a set containing full cross-sectional data of the utility tunnel at multiple time points. Based on the obtained set of full-section data of underground utility tunnels containing multiple time points, a sliding window mechanism is used to perform spatiotemporal slicing of the full-section data of underground utility tunnels to obtain statistical features, time-varying features and spatial features. Specifically, within each spatiotemporal window, statistical features and time-varying features are extracted for the temperature field; spatial features are obtained by calculating the spatial distribution index of gas concentration; and spatiotemporal references of different sensor data are aligned according to the coordinate reference of the spatiotemporal index matrix. Based on the spatiotemporal sliced ​​data, the temporal correlation of feature pairs such as temperature-gas, temperature-smoke, and gas-smoke is quantified using the Pearson correlation coefficient matrix: For each feature pair, the Pearson correlation coefficient of its time series data is calculated to obtain the correlation coefficient matrix, and the matrix element values ​​represent the degree of linear correlation between the corresponding feature pairs; a correlation threshold is set to filter strongly correlated feature pairs with correlation coefficients greater than or equal to the correlation threshold. Based on the selected strongly correlated feature pairs, a spatiotemporal topological connection relationship between features is constructed: each sensor location in the underground utility tunnel is defined as a node in the spatiotemporal topological graph; edges are established between nodes based on the analysis results of temperature gradient changes, smoke diffusion paths, and gas concentration changes to represent the spatiotemporal correlation relationship between features. In the constructed spatiotemporal topology graph, the shortest path length between corresponding nodes of strongly correlated feature pairs is calculated as a quantitative indicator of topological correlation. Combining the correlation coefficient and topological correlation of strongly correlated feature pairs, a dynamic weight is assigned to each edge representing feature correlation. This weight is used to quantify the tightness of the correlation between features, that is, it is determined based on the strong correlation of feature pairs and their connection tightness in the spatiotemporal topology graph. Finally, a multidimensional feature correlation matrix containing three types of correlation relationships: temperature-gas, temperature-smoke, and gas-smoke is generated, where the matrix element values ​​represent the correlation strength of feature pairs, and the correlation strength is determined by the dynamic weight.

[0008] Furthermore, the dynamic rule base construction and optimization module, based on a multi-dimensional feature association matrix, generates candidate itemsets through data transformation and initial pruning. Subsequently, it combines spatiotemporal weighted support to filter frequent itemsets and implements a multi-stage pruning optimization strategy. The specific process of constructing the fire risk rule knowledge base includes: Obtain the generated multidimensional feature association matrix and define a spatiotemporal weight factor θ, which is a dynamic weighted average of the temporal correlation weight α and the spatial proximity weight β, i.e., θ = α + β, α + β = 1. Integrate the spatiotemporal weight factor θ into the support calculation process, introducing spatiotemporally weighted support to optimize the support calculation. S'(X→Y) = [θ × Spatiotemporal Weighted Support (X∪Y) + (1-θ) × Global Support (X∪Y)] / Total Number of Transactions, where S'(X→Y) is the optimized dynamic support; X and Y are features; (X∪Y) represents a feature pair; Spatiotemporal Weighted Support (X∪Y) = Number of Co-occurrences within a Time Window × Spatial Neighborhood Score / Total Number of Transactions, where the number of co-occurrences within a time window represents the number of times the feature pair is triggered together within the sliding time window, and the spatial neighborhood score is calculated based on the spatial distance between sensor nodes, with higher scores for closer distances; Global Support (X∪Y) = Number of Co-occurrences in the Entire Time and Space / Total Number of Transactions; the total number of transactions represents the total number of spatiotemporal windows. In the underground utility tunnel scenario, one transaction corresponds to one spatiotemporal slice; Data processing is performed based on the multidimensional feature correlation matrix to generate a candidate set, and the first pruning operation is performed: the multidimensional feature correlation matrix is ​​converted into transactional data, that is, each spatiotemporal window corresponds to a transaction, and the content of the transaction is a set of strongly correlated feature pairs within the window; Sort feature pairs in descending order of global support; traverse all transactions and record the shared prefix path of feature pairs; convert transaction data to a vertical format to accelerate the frequent itemset counting process; simultaneously, based on the minimum description length (MDL) principle, initially prune low-confidence rules to reduce the size of candidate itemsets. ; Set pruning threshold This is used to determine whether to prune low-confidence rules; if If the conditions are met, then pruning is performed; only candidate sets that satisfy the conditions of spatial proximity and temporal continuity are retained, and Laplace noise is injected during the generation of candidate sets to protect data privacy. By employing association analysis optimization strategies to generate frequent itemsets, and performing a second pruning operation: In the frequent itemset generation stage, an optimized support calculation formula S' is introduced, and an adaptive minimum support threshold is used to further prune the candidate itemsets, removing rules that do not meet the requirements of spatiotemporal weighted support. Based on the confidence threshold Cmin and the lift threshold Lmin, the generated candidate options are subjected to dual filtering: the confidence CL and lift ID of each candidate option are calculated; the confidence CL is calculated as the conditional probability of the rule consequent occurring, that is, the probability of the consequent occurring when the rule antecedent occurs; the lift ID is calculated as the ratio of the probability of the rule consequent and antecedent occurring simultaneously to the probability of the consequent occurring alone, used to measure the degree of increase in the probability of the rule consequent occurring; the confidence threshold Cmin and the lift threshold Lmin are set, and the candidate option is allowed to be retained only if the confidence CL of the candidate option satisfies CL≥Cmin and its lift ID satisfies ID≥Lmin. For candidate itemsets that meet the confidence and lift thresholds, calculate their association rule weights War, where War = Confidence CL × Lift ID × Spatiotemporal weight factor θ. Sort the rules in descending order of their association rule weights and select the top-N high-strength association rules, which are the final frequent itemsets. By selecting the top-N rules, a set of the most representative and important association rules is obtained, which is used to construct the fire risk rule knowledge base. Here, Top-N refers to selecting the top N rules with the highest weights in the ranking of association rule weights War.

[0009] Furthermore, the hierarchical risk warning output module establishes a decision tree model based on a fire risk rule knowledge base, and the process of analyzing the association rule weights of the decision tree branches includes: Extract association rules and their weights from the fire risk rule knowledge base; use the antecedent rule, consequent rule, and corresponding association rule weight of each association rule as features and labels of the data sample; combine all the transformed samples of association rules into a dataset for building a decision tree, where each sample contains feature combination, fire risk label, and association rule weight; Using the CART classification and regression tree algorithm, the association rule weights in the fire risk rule knowledge base are used as the node splitting evaluation index. The decision tree is initialized, and starting from the root node, for each feature, the information gain after splitting according to that feature and splitting point is calculated for all possible splitting points. For each feature A, the information gain at all possible splitting points s is calculated. : In the formula, The dataset of the current node, that is, the set of samples contained in the current node when the node is split; To determine the value of feature A The partitioned subset, i.e., the dataset at the current node In the example, feature A takes the value of The set consisting of all samples; For variables Iterate through all possible values ​​of feature A; For a subset of datasets The number of samples is used to incorporate sample weights when calculating entropy and the number of samples in a subset. ,in For sample index, For the first The association rule weights for each sample; For dataset The entropy, where the formula for calculating entropy is: In the formula, The number of categories for fire risk labels; For fire risk label indexing; For dataset The middle belongs to the first The sample proportion of fire risk labels , For dataset The number of samples, For dataset The middle belongs to the first Number of samples for fire risk labels; Select the feature and split point with the greatest information gain to split the node, and repeat this process until the stopping condition is met to build a decision tree model. The association rule weight range corresponding to different branches in the decision tree is analyzed, and the association rule weight range is divided into three intervals, which correspond to the three-level early warning thresholds of low fire risk, medium fire risk and high fire risk, respectively, thereby establishing a hierarchical decision system including the three-level early warning thresholds; During the decision tree construction process, the information gain ratio index is introduced to optimize the tree branching conditions. The information gain ratio... : In the formula, For splitting information, the calculation formula is: ; When selecting node splitting features and splitting points, the information gain ratios of different features and splitting points are compared, and the feature and splitting point with the largest information gain ratio are selected for splitting to optimize the tree branching conditions.

[0010] Furthermore, the process by which the graded risk early warning output module classifies fire risk levels and generates composite early warning instruction sets includes: The established hierarchical decision-making system is used to determine the fire risk level. At the same time, the probability of fire risk is calculated by combining historical and real-time data. The location of potential hazards is determined based on the spatial coordinates of sensor data. Time series analysis is used to analyze the trend of monitoring data and predict the spread trend of fire hazards. By integrating fire risk probability, hazard location, and spread trend, a composite early warning instruction set is generated. Each composite early warning instruction contains a specific numerical value of fire risk probability, a detailed description of hazard location, and a prediction result of spread trend, realizing a cross-dimensional mapping from single feature to comprehensive risk. The real-time monitoring data stream is input into the hierarchical decision-making system. In the decision tree model, each node has a defined splitting condition, which is determined based on the feature combination and fire risk label of the samples in the dataset when the decision tree is built. The fire risk level of the data is determined by matching the features of the data in the data stream with the conditions of the nodes in the decision tree layer by layer. Starting from the root node, the data features are compared with the splitting conditions of each node in turn until the leaf node is reached, thereby determining the fire risk level corresponding to the data. Based on the determined fire risk level, the corresponding early warning instructions are selected from the composite early warning instruction set. The hierarchical early warning instruction set is output, which includes fire risk probability, hazard location, and dynamic spread prediction.

[0011] Compared with the prior art, the beneficial effect of the present invention is that by collecting multi-source heterogeneous data streams of underground utility tunnels through multimodal sensing devices, key information about the internal environment of the utility tunnels can be obtained comprehensively and in real time. After data preprocessing, a standardized three-dimensional time-series dataset is generated, providing a high-quality data foundation for subsequent analysis. Using a sliding window spatiotemporal slicing technique, statistical, time-varying, and spatial features are extracted to comprehensively capture the temporal and spatial variations of the utility tunnel's internal environment. Pearson correlation coefficients are used to screen strongly correlated feature pairs, and dynamic weights are employed to quantify the correlation strength, constructing a multi-dimensional feature correlation matrix. This helps to deeply understand the intrinsic connections between features within the utility tunnel's internal environment, providing strong feature support for fire risk prediction. By introducing spatiotemporal weighted support optimization of the Apriori algorithm, candidate itemsets and frequent itemsets are generated, and pruning optimization strategies are implemented to construct a fire risk rule knowledge base. This efficiently uncovers key correlation rules for utility tunnel fire risks, while considering the impact of time and space factors on the rules, making the generated rules more practical and accurate. By establishing a decision tree model, analyzing the correlation rule weights of the decision tree branches, fire risk levels are classified, and a composite early warning instruction set is generated. This transforms complex fire risk information into intuitive risk levels and early warning instructions, facilitating rapid understanding and response by management personnel. Attached Figure Description

[0012] Figure 1 This is a block diagram of the underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm proposed in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 A fire feature correlation analysis system for underground utility tunnels based on an improved Apriori algorithm. The system includes a multimodal data acquisition and preprocessing module, a multidimensional feature correlation matrix construction module, a dynamic rule base construction and optimization module, and a hierarchical risk early warning output module. Multimodal data acquisition and preprocessing module: Collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices, and generates standardized three-dimensional time-series datasets after data preprocessing operations; the multimodal sensing devices include distributed fiber optic temperature sensors, gas sensor arrays, and smoke concentration detection units; Multidimensional feature association matrix construction module: Based on a three-dimensional time series dataset, statistical features, time-varying features, and spatial features are extracted through a sliding window spatiotemporal slice. Strongly correlated feature pairs are selected by combining the Pearson correlation coefficient, and the correlation strength of feature pairs is quantified by dynamic weights to construct a multidimensional feature association matrix. Dynamic rule base construction and optimization module: Based on the multidimensional feature association matrix, candidate itemsets are generated through data transformation and initial pruning. Then, frequent itemsets are screened by combining spatiotemporal weighted support, and a multi-stage pruning optimization strategy is implemented to build a fire risk rule knowledge base. The graded risk warning output module establishes a decision tree model based on the fire risk rule knowledge base, analyzes the association rule weights of the decision tree branches, classifies fire risk levels, and generates a composite warning instruction set.

[0015] It should be further explained that, in the specific implementation process, the multimodal data acquisition and preprocessing module collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices. The process of generating a standardized three-dimensional time-series dataset through data preprocessing includes: Distributed fiber optic temperature sensors, gas sensor arrays, and smoke concentration detection units are deployed in the underground utility tunnel. The distributed fiber optic temperature sensors can be laid along the tunnel to continuously monitor temperature. The gas sensor array contains multiple gas sensors to detect different types of gas concentrations. The smoke concentration detection unit monitors smoke density in real time. The CAN bus protocol is used to achieve synchronous data acquisition from multiple nodes, generating a multi-source heterogeneous data stream of the underground utility tunnel containing temperature, gas concentration, and smoke density. This data stream is the raw data stream. It is understood that the CAN bus protocol has the characteristics of high real-time performance, high reliability, and strong anti-interference capability, ensuring that the data collected by each sensor node remains synchronized in time. To eliminate high-frequency noise from the original data stream, wavelet transform is performed as follows: A wavelet basis function (such as the Daubechies wavelet) is selected, and the number of wavelet decomposition levels is determined based on the characteristics of the original data. Multi-scale wavelet decomposition is then performed on the original data, and the high-frequency noise components (corresponding to the high-frequency coefficients of the wavelet decomposition) are either set to zero or thresholded. Wavelet reconstruction is then performed to obtain the data after high-frequency noise elimination. Moving average filtering is then used to suppress low-frequency drift: The size of the moving average window is set (e.g., 5 data points per window). The data within each window is averaged, and the average value replaces the data point at the center of the window. By continuously sliding the window, the entire data stream is processed, thereby suppressing low-frequency drift. Z-score normalization is used to unify the dimensions of multi-source sensor data. Based on GPS timing and UWB positioning technology, each set of sensor data is assigned a timestamp and spatial coordinates. The GPS timing system provides accurate time information, and UWB positioning technology utilizes ultra-wideband signals to achieve high-precision spatial positioning. The timestamps and spatial coordinates are then associated with the corresponding sensor data. Construct a spatiotemporal index matrix containing timestamps and spatial coordinates: Treat each sensor data as an element of the matrix, arrange them according to time order and spatial location to form a matrix structure, where the rows of the matrix represent the time dimension and the columns represent the spatial dimension. The matrix elements contain sensor data at the corresponding time point and spatial location, thus forming a standardized three-dimensional time-series dataset.

[0016] It should be further explained that, in the specific implementation process, the multidimensional feature association matrix construction module, based on a three-dimensional time-series dataset, extracts statistical features, time-varying features, and spatial features through a sliding window spatiotemporal slice. It then uses the Pearson correlation coefficient to screen strongly correlated feature pairs and utilizes dynamic weights to quantify the correlation strength of feature pairs. The process of constructing the multidimensional feature association matrix includes: In the constructed 3D time-series dataset, data from different spatial locations (i.e., sensor locations) at the same timestamp are integrated based on the spatial layout of the underground utility tunnel. Specifically, sensor data from the tunnel cross-section are categorized and summarized according to their spatial coordinates. For example, sensor data such as temperature, gas concentration, and smoke density at different heights and angles of a tunnel cross-section are combined to form a complete data set for that cross-section at a specific time point, i.e., the full cross-sectional data of the utility tunnel corresponding to that time point. By traversing all time points in the 3D time-series dataset, the full cross-sectional data of the utility tunnel corresponding to each time point is generated sequentially, thus obtaining a set containing full cross-sectional data of the utility tunnel at multiple time points, providing basic data for the subsequent multi-dimensional feature association matrix construction module. Based on the obtained set of full-section data of underground utility tunnels containing multiple time points, a sliding window mechanism is used to spatiotemporally slice the data and obtain statistical, time-varying, and spatial features. The sliding window moves in both time and space dimensions: in the time dimension, the window slides forward according to a preset time step, and after each slide, the window contains full-section data of the utility tunnel within a certain time period; in the spatial dimension, the sliding window is determined to cover part or all of the cross-sectional area of ​​the utility tunnel. For example, the size of the sliding window is determined, such as a time window length of 1 hour and a spatial window range of 10 meters, and a sliding step size is set, such as a time step size of 30 minutes and a spatial step size of 5. The sliding window, according to the set time and spatial steps, divides the data into slices with spatiotemporal attributes. Each slice represents a specific spatiotemporal window, containing the underground utility tunnel data at the corresponding time and spatial location within that window. The system uses full-section data of the utility tunnel. Within each spatiotemporal window, statistical and time-varying features of the temperature field are extracted. For statistical features, the mean of the temperature field within the spatiotemporal window is calculated (i.e., summing all temperature data within the window and dividing by the number of data points). The variance is calculated by averaging the squares of the differences between each data point and the mean. The extreme values ​​within the window, i.e., the maximum and minimum values, are identified. For time-varying features, the temperature time series data is converted to the frequency domain using Fourier transform, and the fluctuation frequency is calculated based on the frequency domain information. The temperature time series is linearly fitted using the least squares method, and the slope of the fitted line is the trend slope. Spatial features are obtained by calculating the spatial distribution index of gas concentration: for the spatial coefficient of variation, the mean and standard deviation of the gas concentration at each spatial location are calculated, and the spatial coefficient of variation is the ratio of the standard deviation to the mean. For local clustering, spatial autocorrelation analysis methods (such as Moran's method) are used. The I-index measures local aggregation by comparing the similarity of gas concentration at each spatial location with that at its neighboring locations. For diffusion rate, the diffusion distance per unit time is calculated based on the spatial distribution of gas concentration at different time points, combined with spatial distance. The spatiotemporal reference of the spatiotemporal index matrix is ​​used to align the data from different sensors. This is understandable because different sensors may have measurement errors or acquisition time deviations; the spatiotemporal index matrix unifies the data from each sensor to the same time and space coordinate system, ensuring the accuracy of subsequent feature correlation analysis. These spatiotemporally sliced ​​data provide rich spatiotemporal samples for subsequent quantification of the temporal correlation of feature pairs, making the analysis results more accurate and reliable. Based on the spatiotemporal sliced ​​data, the temporal correlation of feature pairs such as temperature-gas, temperature-smoke, and gas-smoke is quantified using the Pearson correlation coefficient matrix: For each feature pair, the Pearson correlation coefficient of its time series data is calculated to obtain a correlation coefficient matrix, where the matrix element values ​​represent the degree of linear correlation between the corresponding feature pairs; a correlation threshold is set to filter strongly correlated feature pairs (such as temperature-gas and temperature-smoke combinations) with correlation coefficients greater than or equal to the threshold; it is understandable that, in addition to the pairwise combinations of temperature-gas, temperature-smoke, and gas-smoke feature pairs, combinations involving three features (i.e., feature triples) can be used for more complex and diverse scenarios; Based on the selected strongly correlated feature pairs, a spatiotemporal topological connection relationship between features is constructed: each sensor location in the underground utility tunnel (including distributed fiber optic temperature sensors, smoke concentration detection units, and gas sensor arrays) is defined as a node in the spatiotemporal topological graph; edges are established between nodes based on the analysis results of temperature gradient changes, smoke diffusion paths, and gas concentration changes to represent the spatiotemporal correlation between features. The specific analysis process is as follows: time series analysis is performed on the temperature data collected by the distributed fiber optic temperature sensors to identify the temperature change trend at different time points; combined with the spatial layout of the underground utility tunnel, the temperature distribution differences at different spatial locations are analyzed to form a temperature gradient map; by analyzing the changes in temperature gradient at different times and spatial locations, the causal relationship between temperature changes and air flow and smoke diffusion is explored. For example, an increase in temperature gradient may lead to accelerated air flow, thereby affecting the smoke diffusion path; the smoke density in the underground utility tunnel is monitored in real time using the smoke concentration detection unit; based on the spatial distribution and temporal changes of smoke concentration, the diffusion path of smoke in the utility tunnel, i.e., direction and speed, is simulated; the concentration of different types of gases in the utility tunnel is monitored in real time using the gas sensor array; the trend of gas concentration changes over time is analyzed to identify abnormal fluctuations in gas concentration. In the constructed spatiotemporal topology graph, the shortest path length between corresponding nodes of strongly correlated feature pairs is calculated as a quantitative indicator of topological correlation. The shortest path length reflects the tightness of connection between two nodes in the spatiotemporal topology graph; the shorter the path, the more direct and tighter the connection between the two nodes. If the temperature-gas feature pair is strongly correlated, the shortest path length between the distributed fiber optic temperature sensor node and the gas sensor array node is calculated, and this length value serves as an indicator of the topological correlation between these two features. Combining the correlation coefficient and topological correlation of strongly correlated feature pairs, a dynamic weight is assigned to each edge representing feature correlation. This weight is used to quantify the tightness of the connection between features, i.e., determined based on the strong correlation of feature pairs and their tightness of connection in the spatiotemporal topology graph. Specifically, the correlation coefficient of strongly correlated feature pairs is calculated. Assign a basic relevance weight. This weight is a relation to A function with positive correlation to the value ,in It is an increasing function; for each feature pair, it is calculated based on the shortest path length between their corresponding nodes. Assign a topological association weight This weight is a relation to Functions with negative correlation values ,in It is a decreasing function; combined with correlation weights Weights associated with topology The final dynamic weights are determined using a weighted average method. In the formula, For dynamic weights, For dynamic adjustment coefficients, and Finally, a multidimensional feature association matrix is ​​generated, which includes three types of associations: temperature-gas, temperature-smoke, and gas-smoke. The matrix element values ​​represent the association strength of feature pairs, and the association strength is determined by dynamic weights.

[0017] It should be further explained that, in the specific implementation process, the dynamic rule base construction and optimization module, based on a multi-dimensional feature association matrix, generates candidate itemsets through data transformation and initial pruning. Subsequently, it combines spatiotemporal weighted support to filter frequent itemsets and implements a multi-stage pruning and optimization strategy. The specific process of constructing the fire risk rule knowledge base includes: The generated multidimensional feature correlation matrix is ​​obtained, and a spatiotemporal weighting factor θ is defined, which is a dynamic weighting of the temporal correlation weight α and the spatial proximity weight β, i.e., θ = α + β, α + β = 1. The temporal correlation weight α is determined based on the strength of the temporal correlation between feature pairs; the larger the Pearson correlation coefficient, the larger α. The spatial proximity weight β is determined based on the spatial distance between sensor nodes and the spatial correlation between features; the closer the distance and the stronger the spatial correlation, the larger β. The spatiotemporal weighting factor θ is integrated into the support calculation process, introducing spatiotemporally weighted support to optimize the support calculation. S'(X→Y) = [θ × Spatiotemporal Weighted Support (X∪Y) + (1-θ) × Global Support (X∪Y)] / Total Number of Transactions, where S'(X→Y) is the optimized dynamic support; X and Y are features; (X∪Y) represents a feature pair; Spatiotemporal Weighted Support (X∪Y) = Co-occurrence Count within Time Window × Spatial Neighborhood Score / Total Number of Transactions, where the co-occurrence count within the time window represents the number of times feature pairs (such as temperature and gas concentration) are triggered together within the sliding time window, and the spatial neighborhood score is calculated based on the spatial distance between sensor nodes, with higher scores for closer distances; Global Support (X∪Y) = Co-occurrence Count across All Time Periods and Space / Total Number of Transactions; Total Number of Transactions Table The total number of spatiotemporal windows is shown. In the underground utility tunnel scenario, one transaction corresponds to one spatiotemporal slice. It can be understood that the spatiotemporal weighted support is a support metric calculated by combining the number of co-occurrences of feature pairs within a specific time window and the spatial proximity score of sensor nodes, and then dividing this ratio by the total number of transactions. It reflects the temporal and spatial correlation of feature pairs. Global support refers to the ratio of the number of times feature pairs co-occur throughout the entire dataset (covering all time periods and spaces) to the total number of transactions, used to measure the overall correlation of feature pairs. By introducing spatiotemporal weight factors and spatiotemporal weighted support, the support calculation process is optimized, making the calculation results more consistent with the actual correlation of underground utility tunnel fire features. Data processing is performed based on the multidimensional feature correlation matrix to generate a candidate set, followed by the first pruning operation: the multidimensional feature correlation matrix is ​​converted into transactional data, that is, each spatiotemporal window corresponds to one transaction, and the transaction content is a set of strongly correlated feature pairs within the window (such as {abnormal temperature, excessive gas}); a small number of non-correlated feature pairs (such as {normal temperature, normal gas}) are randomly generated as negative samples to increase the generalization ability of the model. Sort feature pairs in descending order of global support (e.g., {temperature anomaly, gas exceedance} → {temperature anomaly, smoke exceedance}); traverse all transactions, recording the shared prefix path of feature pairs. For example, transactions {B,C} and {B,D} share the prefix B, which is merged into B→C / D; convert transaction data into a vertical format (feature pair, spatiotemporal window identifier) ​​to accelerate frequent itemset counting; simultaneously, based on the minimum description length (MDL) principle, initially prune low-confidence rules to reduce the size of candidate itemsets. ; Set pruning threshold This is used to determine whether to prune low-confidence rules; if If the conditions are not met, pruning is performed; only candidate sets that meet the conditions of spatial proximity (sensor spacing ≤ 10m) and temporal continuity (time window overlap ≥ 50%) are retained; Laplace noise is injected during the generation of candidate sets to protect data privacy. In the formula, Count the support of the perturbed feature pairs; This is the original support count for the feature pairs; It follows a Laplace distribution, where 0 is the mean of the distribution, indicating that the expected value of the noise is zero, ensuring that the noise does not systematically deviate from the true value; The scale parameter of the distribution (to control noise intensity); For sensitivity, its value depends on the range of changes in the count of feature pairs. For example, if the number of occurrences of a feature pair changes at most once due to the addition or subtraction of a single transaction, then... =1; The core parameter for differential privacy represents the strictness of the algorithm's privacy protection. It is understandable that by sorting global support and pruning MDL, a preliminary optimized candidate set is quickly generated, providing a foundation for further analysis. A frequent itemset generation strategy is employed using association analysis to generate frequent itemsets, followed by a second pruning operation: During the frequent itemset generation stage, an optimized support calculation formula S' is introduced, and an adaptive minimum support threshold (based on iterative optimization using historical warning data) is used to further prune the candidate itemsets, removing rules that do not meet the spatiotemporal weighted support requirements. The generated candidate itemsets are then subjected to dual filtering based on a confidence threshold Cmin and a lift threshold Lmin: the confidence CL and lift ID of each candidate itemset (i.e., potential association rules) are calculated. The confidence CL is calculated as the conditional probability of the rule consequent occurring, i.e., the probability of the consequent occurring when the rule antecedent occurs. The lift ID is calculated as the ratio of the probability of the rule consequent and antecedent occurring simultaneously to the probability of the consequent occurring alone, used to measure the degree of increase in the probability of the rule consequent occurring. A confidence threshold Cmin and a lift threshold Lmin are set, and a candidate itemset is allowed to be retained only if its confidence CL satisfies CL≥Cmin and its lift ID satisfies ID≥Lmin. For candidate itemsets that meet the confidence and lift thresholds, their association rule weights (War) are calculated, where War = Confidence CL × Lift ID × Spatiotemporal weight factor θ. Rules are sorted in descending order of association rule weights, and the top-N high-strength association rules are selected. These rules constitute the final frequent itemsets (including antecedent-consequence rules and association rule weights). By selecting the top-N rules, a set of the most representative and important association rules is obtained, which is used to construct the fire risk rule knowledge base. Here, Top-N refers to selecting the top N rules with the highest weights in the association rule weight War ranking, where N is a preset value representing the number of association rules to be retained. It is understood that through the second pruning operation and association rule weight calculation, high-quality and highly reliable association rules are generated, providing strong support for fire risk prediction.

[0018] It should be further explained that, in the specific implementation process, the hierarchical risk early warning output module establishes a decision tree model based on the fire risk rule knowledge base, analyzes the association rule weights of the decision tree branches, classifies fire risk levels, and generates a composite early warning instruction set as follows: The association rules and their weights are extracted from the fire risk rule knowledge base. The antecedent rule, consequent rule, and corresponding association rule weight of each association rule are used as features and labels for the data sample. For example, an association rule "temperature rises and gas concentration exceeds the standard → high fire risk, association rule weight is 0.8", with the antecedent "temperature rises and gas concentration exceeds the standard" as the feature combination, the consequent "high fire risk" as the label, and the association rule weight of 0.8 as the weight value of the sample. All samples transformed from the association rules are used to form a dataset for building a decision tree. Each sample in the dataset contains a feature combination, a fire risk label, and an association rule weight. Here, "temperature rises and gas concentration exceeds the standard" refers to the joint condition of multiple features in the antecedent of the rule (e.g., "temperature rises and gas concentration exceeds the standard"), which is an extended form of feature pairs. A feature combination consists of one or more feature pairs; for example, the antecedent "temperature rises and gas concentration exceeds the standard" may implicitly contain multiple low-level feature pairs such as "temperature-gas", "temperature-time", and "gas-space". Using the CART classification and regression tree algorithm, the association rule weights in the fire risk rule knowledge base are used as the node splitting evaluation index. The decision tree is initialized, and starting from the root node, for each feature, the information gain after splitting according to that feature and splitting point is calculated for all possible splitting points. For each feature A, the information gain at all possible splitting points s is calculated. : In the formula, The dataset of the current node, that is, the set of samples contained in the current node when the node is split; To determine the value of feature A The partitioned subset, i.e., the dataset at the current node In the example, feature A takes the value of The set consisting of all samples; For variables Iterate through all possible values ​​of feature A; For a subset of datasets The number of samples is used to incorporate sample weights when calculating entropy and the number of samples in a subset. ,in For sample index, For the first The association rule weights for each sample; For dataset The entropy, where the formula for calculating entropy is: In the formula, The number of categories for fire risk labels (e.g., low fire risk, medium fire risk, and high fire risk). For fire risk label indexing; For dataset The middle belongs to the first The sample proportion of fire risk labels , For dataset The number of samples, For dataset The middle belongs to the first Number of samples for fire risk labels; Select the feature and split point with the largest information gain to split the node, and repeat this process until the stopping condition is met (i.e., the number of samples in the node is less than the threshold, the information gain is less than the threshold, or the depth of the tree reaches the preset value), and build a decision tree model. The association rule weight range corresponding to different branches in the decision tree is analyzed and divided into three intervals, which correspond to the three-level early warning thresholds of low fire risk, medium fire risk and high fire risk. For example, the association rule weight is set to be less than 0.3 as the early warning threshold for low fire risk, between 0.3 and 0.6 as the early warning threshold for medium fire risk, and greater than 0.6 as the early warning threshold for high fire risk, thereby establishing a hierarchical decision system that includes three-level early warning thresholds. During the decision tree construction process, the information gain ratio index is introduced to optimize the tree branching conditions. The information gain ratio... : In the formula, For splitting information, the calculation formula is: ; When selecting node splitting features and splitting points, compare the information gain ratios of different features and splitting points, and select the feature and splitting point with the highest information gain ratio for splitting. This avoids information gain bias towards features with more values ​​and optimizes the tree branching conditions. The established hierarchical decision-making system is used to determine the fire risk level. Simultaneously, combining historical and real-time data, the fire risk probability is calculated. This involves estimating the fire risk probability under the current characteristic combination based on the ratio of the number of fires occurring under the same characteristic combination over a past period to the total number of monitoring sessions. The spatial coordinates of sensor data are used to determine the location of potential hazards. This is achieved by using the spatial coordinates of distributed fiber optic temperature sensors, gas sensor arrays, and smoke concentration detection units to pinpoint the specific locations of temperature anomalies, excessive gas concentrations, or smoke occurrences. Time series analysis methods (such as moving averages and exponential smoothing) are used to perform trend analysis on the monitoring data to predict the spread trend of fire hazards. This involves predicting the direction and speed of change of these parameters over a future period based on changes in temperature, gas concentration, and smoke density over a past period, thereby obtaining the spread trend. By integrating fire risk probability, hazard location, and spread trend, a composite early warning instruction set is generated. Each composite early warning instruction includes a specific numerical value of fire risk probability, a detailed description of hazard location (specific section and coordinates of underground utility tunnel), and a predicted spread trend (spread direction, speed, etc.), realizing a cross-dimensional mapping from single features to comprehensive risks. The real-time monitoring data stream is input into the hierarchical decision-making system. In the decision tree model, each node defines corresponding splitting conditions, which are determined based on the feature combination of the central samples in the dataset and the fire risk label when the decision tree is built. The fire risk level of the data is determined by matching the features of the data in the data stream with the conditions of the nodes in the decision tree layer by layer. Starting from the root node, the data features are compared with the splitting conditions of each node in turn until the leaf node is reached, thereby determining the fire risk level corresponding to the data. Based on the determined fire risk level, the corresponding early warning instructions are selected from the composite early warning instruction set. The hierarchical early warning instruction set is output, which includes fire risk probability, hazard location, and dynamic spread prediction. For example, for a high fire risk warning, the output is a high fire risk probability, a specific hazard location, and a prediction of rapid spread; for a low fire risk warning, the output is a low fire risk probability, a general hazard location, and a prediction of slow or no spread.

[0019] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0020] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0021] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0022] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fire feature correlation analysis system for underground utility tunnels based on an improved Apriori algorithm, characterized in that: It includes a multimodal data acquisition and preprocessing module, a multidimensional feature association matrix construction module, a dynamic rule base construction and optimization module, and a hierarchical risk warning output module; Multimodal data acquisition and preprocessing module: Collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices, and generates standardized three-dimensional time-series datasets after data preprocessing. Multidimensional feature association matrix construction module: Based on a three-dimensional time series dataset, statistical features, time-varying features, and spatial features are extracted through a sliding window spatiotemporal slice. Strongly correlated feature pairs are selected by combining the Pearson correlation coefficient, and the correlation strength of feature pairs is quantified by dynamic weights to construct a multidimensional feature association matrix. Dynamic rule base construction and optimization module: Based on the multidimensional feature association matrix, candidate itemsets are generated through data transformation and initial pruning. Then, frequent itemsets are screened by combining spatiotemporal weighted support, and a multi-stage pruning optimization strategy is implemented to build a fire risk rule knowledge base. The graded risk warning output module establishes a decision tree model based on the fire risk rule knowledge base, analyzes the association rule weights of the decision tree branches, classifies fire risk levels, and generates a composite warning instruction set.

2. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: In the multimodal data acquisition and preprocessing module, the multimodal sensing devices include a distributed fiber optic temperature sensor, a gas sensor array, and a smoke concentration detection unit.

3. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: The multimodal data acquisition and preprocessing module collects multi-source heterogeneous data streams from underground utility tunnels through multimodal sensing devices. The process of generating a standardized 3D time-series dataset through data preprocessing includes: Distributed fiber optic temperature sensors, gas sensor arrays, and smoke concentration detection units are deployed in the underground utility tunnel; multi-node data synchronous acquisition is achieved using the CAN bus protocol to generate a multi-source heterogeneous data stream of the underground utility tunnel containing temperature, gas concentration, and smoke density. This data stream is the raw data stream. Wavelet transform is applied to the raw data stream to eliminate high-frequency noise, moving average filtering is used to suppress low-frequency drift, and Z-score normalization is used to unify the dimensions of multi-source sensor data; based on GPS timing system and UWB positioning technology, each set of sensor data is assigned a timestamp and spatial coordinates. Construct a spatiotemporal index matrix containing timestamps and spatial coordinates: Treat each sensor data as an element of the matrix, arrange them according to time order and spatial location to form a matrix structure, where the rows of the matrix represent the time dimension and the columns represent the spatial dimension. The matrix elements contain sensor data at the corresponding time point and spatial location, thus forming a standardized three-dimensional time-series dataset.

4. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: The multidimensional feature association matrix construction module, based on a three-dimensional time-series dataset, extracts statistical, time-varying, and spatial features through a sliding window spatiotemporal slice. It then uses the Pearson correlation coefficient to filter strongly correlated feature pairs and employs dynamic weights to quantify the association strength of these pairs. The process of constructing the multidimensional feature association matrix includes: In the constructed three-dimensional time series dataset, based on the spatial layout of the underground utility tunnel, data from different spatial locations at the same timestamp are integrated: sensor data on the cross-section of the utility tunnel are classified and summarized according to their spatial coordinates; by traversing all time points in the three-dimensional time series dataset, the full cross-sectional data of the utility tunnel corresponding to each time point is generated in sequence, thus obtaining a set containing full cross-sectional data of the utility tunnel at multiple time points. Based on the obtained set of full-section data of underground utility tunnels containing multiple time points, a sliding window mechanism is used to perform spatiotemporal slicing of the full-section data of underground utility tunnels to obtain statistical features, time-varying features and spatial features. Specifically, within each spatiotemporal window, statistical features and time-varying features are extracted for the temperature field; spatial features are obtained by calculating the spatial distribution index of gas concentration; and spatiotemporal references of different sensor data are aligned according to the coordinate reference of the spatiotemporal index matrix. Based on the spatiotemporal sliced ​​data, the temporal correlation of feature pairs such as temperature-gas, temperature-smoke, and gas-smoke is quantified using the Pearson correlation coefficient matrix: For each feature pair, the Pearson correlation coefficient of its time series data is calculated to obtain the correlation coefficient matrix, and the matrix element values ​​represent the degree of linear correlation between the corresponding feature pairs; a correlation threshold is set to filter strongly correlated feature pairs with correlation coefficients greater than or equal to the correlation threshold. Based on the selected strongly correlated feature pairs, a spatiotemporal topological connection relationship between features is constructed: each sensor location in the underground utility tunnel is defined as a node in the spatiotemporal topological graph; edges are established between nodes based on the analysis results of temperature gradient changes, smoke diffusion paths, and gas concentration changes to represent the spatiotemporal correlation relationship between features. In the constructed spatiotemporal topology graph, the shortest path length between corresponding nodes of strongly correlated feature pairs is calculated as a quantitative indicator of topological correlation. Combining the correlation coefficient and topological correlation of strongly correlated feature pairs, a dynamic weight is assigned to each edge representing feature correlation. This weight is used to quantify the tightness of the correlation between features, that is, it is determined based on the strong correlation of feature pairs and their connection tightness in the spatiotemporal topology graph. Finally, a multidimensional feature correlation matrix containing three types of correlation relationships: temperature-gas, temperature-smoke, and gas-smoke is generated, where the matrix element values ​​represent the correlation strength of feature pairs, and the correlation strength is determined by the dynamic weight.

5. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: The dynamic rule base construction and optimization module, based on a multi-dimensional feature association matrix, generates candidate itemsets through data transformation and initial pruning. Subsequently, it filters frequent itemsets using spatiotemporal weighted support and implements a multi-stage pruning and optimization strategy. The specific process of constructing the fire risk rule knowledge base includes: Obtain the generated multidimensional feature association matrix and define a spatiotemporal weight factor θ, which is a dynamic weighted average of the temporal correlation weight α and the spatial proximity weight β, i.e., θ = α + β, α + β = 1. Integrate the spatiotemporal weight factor θ into the support calculation process, introducing spatiotemporally weighted support to optimize the support calculation. S'(X→Y)=[θ×spatiotemporal weighted support(X∪Y)+(1-θ)×global support(X∪Y)] / total number of transactions, In the formula, S'(X→Y) is the optimized dynamic support; X and Y are features; (X∪Y) represents a feature pair; the spatiotemporal weighted support (X∪Y) = number of co-occurrences within the time window × spatial neighborhood score / total number of transactions, where the number of co-occurrences within the time window represents the number of times the feature pair is triggered together within the sliding time window, and the spatial neighborhood score is calculated based on the spatial distance between sensor nodes, with higher scores for closer distances; the global support (X∪Y) = number of co-occurrences across all time periods and spaces / total number of transactions; the total number of transactions represents the total number of spatiotemporal windows, and in the underground utility tunnel scenario, one transaction corresponds to one spatiotemporal slice; Data processing is performed based on the multidimensional feature correlation matrix to generate a candidate set, and the first pruning operation is performed: the multidimensional feature correlation matrix is ​​converted into transactional data, that is, each spatiotemporal window corresponds to a transaction, and the content of the transaction is a set of strongly correlated feature pairs within the window; Sort feature pairs in descending order of global support; traverse all transactions and record the shared prefix path of feature pairs; convert transaction data to a vertical format to accelerate the frequent itemset counting process; simultaneously, based on the minimum description length (MDL) principle, initially prune low-confidence rules to reduce the size of candidate itemsets. ； Set pruning threshold This is used to determine whether to prune low-confidence rules; if If the conditions are met, then pruning is performed; only candidate sets that satisfy the conditions of spatial proximity and temporal continuity are retained, and Laplace noise is injected during the generation of candidate sets to protect data privacy. By employing association analysis optimization strategies to generate frequent itemsets, and performing a second pruning operation: In the frequent itemset generation stage, an optimized support calculation formula S' is introduced, and an adaptive minimum support threshold is used to further prune the candidate itemsets, removing rules that do not meet the requirements of spatiotemporal weighted support. Based on the confidence threshold Cmin and the lift threshold Lmin, the generated candidate options are subjected to dual filtering: the confidence CL and lift ID of each candidate option are calculated; the confidence CL is calculated as the conditional probability of the rule consequent occurring, that is, the probability of the consequent occurring when the rule antecedent occurs; the lift ID is calculated as the ratio of the probability of the rule consequent and antecedent occurring simultaneously to the probability of the consequent occurring alone, used to measure the degree of increase in the probability of the rule consequent occurring; the confidence threshold Cmin and the lift threshold Lmin are set, and the candidate option is allowed to be retained only if the confidence CL of the candidate option satisfies CL≥Cmin and its lift ID satisfies ID≥Lmin. For candidate itemsets that meet the confidence and lift thresholds, calculate their association rule weights War, where War = Confidence CL × Lift ID × Spatiotemporal weight factor θ. Sort the rules in descending order of their association rule weights and select the top-N high-strength association rules, which are the final frequent itemsets. By selecting the top-N rules, a set of the most representative and important association rules is obtained, which is used to construct the fire risk rule knowledge base. Here, Top-N refers to selecting the top N rules with the highest weights in the ranking of association rule weights War.

6. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: The graded risk early warning output module establishes a decision tree model based on a fire risk rule knowledge base. The process of analyzing the association rule weights of the decision tree branches includes: Extract association rules and their weights from the fire risk rule knowledge base; use the antecedent rule, consequent rule, and corresponding association rule weight of each association rule as features and labels of the data sample; combine all the transformed samples of association rules into a dataset for building a decision tree, where each sample contains feature combination, fire risk label, and association rule weight; Using the CART classification and regression tree algorithm, the association rule weights in the fire risk rule knowledge base are used as the node splitting evaluation index. The decision tree is initialized, and starting from the root node, for each feature, the information gain after splitting according to that feature and splitting point is calculated for all possible splitting points. For each feature A, the information gain at all possible splitting points s is calculated. : In the formula, The dataset of the current node, that is, the set of samples contained in the current node when the node is split; To determine the value of feature A The partitioned subset, i.e., the dataset at the current node In the example, feature A takes the value of The set consisting of all samples; For variables Iterate through all possible values ​​of feature A; For a subset of datasets The number of samples is used to incorporate sample weights when calculating entropy and the number of samples in a subset. ,in For sample index, For the first The association rule weights for each sample; For dataset The entropy, where the formula for calculating entropy is: In the formula, The number of categories for fire risk labels; For fire risk label indexing; For dataset The middle belongs to the first The sample proportion of fire risk labels , For dataset The number of samples, For dataset The middle belongs to the first Number of samples for fire risk labels; Select the feature and split point with the greatest information gain to split the node, and repeat this process until the stopping condition is met to build a decision tree model. The association rule weight range corresponding to different branches in the decision tree is analyzed, and the association rule weight range is divided into three intervals, which correspond to the three-level early warning thresholds of low fire risk, medium fire risk and high fire risk, respectively, thereby establishing a hierarchical decision system including the three-level early warning thresholds; During the decision tree construction process, the information gain ratio index is introduced to optimize the tree branching conditions. The information gain ratio... : In the formula, For splitting information, the calculation formula is: ； When selecting node splitting features and splitting points, the information gain ratios of different features and splitting points are compared, and the feature and splitting point with the largest information gain ratio are selected for splitting to optimize the tree branching conditions.

7. The underground utility tunnel fire feature association analysis system based on the improved Apriori algorithm according to claim 1, characterized in that: The process by which the graded risk early warning output module classifies fire risk levels and generates composite early warning instruction sets includes: The established hierarchical decision-making system is used to determine the fire risk level. At the same time, the probability of fire risk is calculated by combining historical and real-time data. The location of potential hazards is determined based on the spatial coordinates of sensor data. Time series analysis is used to analyze the trend of monitoring data and predict the spread trend of fire hazards. By integrating fire risk probability, hazard location, and spread trend, a composite early warning instruction set is generated. Each composite early warning instruction contains a specific numerical value of fire risk probability, a detailed description of hazard location, and a prediction result of spread trend, realizing a cross-dimensional mapping from single feature to comprehensive risk. The real-time monitoring data stream is input into the hierarchical decision-making system. In the decision tree model, each node has a defined splitting condition, which is determined based on the feature combination and fire risk label of the samples in the dataset when the decision tree is built. The fire risk level of the data is determined by matching the features of the data in the data stream with the conditions of the nodes in the decision tree layer by layer. Starting from the root node, the data features are compared with the splitting conditions of each node in turn until the leaf node is reached, thereby determining the fire risk level corresponding to the data. Based on the determined fire risk level, the corresponding early warning instructions are selected from the composite early warning instruction set. The hierarchical early warning instruction set is output, which includes fire risk probability, hazard location, and dynamic spread prediction.

Images